Embedded analytics of animal images

• Main Author: Thomassen, Sigurd
• Format: Master Thesis
• Language: English
• Published: UiT Norges arktiske universitet 2017
• Subjects: VDP::Matematikk og Naturvitenskap: 400::Informasjons- og kommunikasjonsvitenskap: 420::Kommunikasjon og distribuerte systemer: 423; VDP::Mathematics and natural science: 400::Information and communication science: 420::Communication and distributed systems: 423; VDP::Matematikk og Naturvitenskap: 400::Informasjons- og kommunikasjonsvitenskap: 420::Kunnskapsbaserte systemer: 425; VDP::Mathematics and natural science: 400::Information and communication science: 420::Knowledge based systems: 425; VDP::Matematikk og Naturvitenskap: 400::Informasjons- og kommunikasjonsvitenskap: 420::Systemutvikling og – arbeid: 426; VDP::Mathematics and natural science: 400::Information and communication science: 420::System development and system design: 426; VDP::Matematikk og Naturvitenskap: 400::Informasjons- og kommunikasjonsvitenskap: 420::Simulering; visualisering; signalbehandling; bildeanalyse: 429; VDP::Mathematics and natural science: 400::Information and communication science: 420::Simulation; visualization; signal processing; image processing: 429; INF-3981; Arctic; Norway; Climate change; Finnmark; Tundra
• Online Access: https://hdl.handle.net/10037/12000

Due to the large increase of image data in animal surveillance, an effective and efficient way of labeling said data is required. Over the past few years the Climate-ecological Observatory for Arctic Tundra (COAT) project have deployed dozens of cameras in eastern Finnmark, Norway during winter, which have resulted in a large volume of wildlife images which is used to document the effects of climate change on animal ecosystems in the area. The images are manually labeled by biologists, and is a time-consuming task. This thesis presents the architecture, design and implementation of an image classification system to be used with the camera traps for in-situ analytics on accumulated image data for periodical updates. The system will automatically classify and label the images taken by the cameras. Using state-of-the-art Convolutional Neural Networks (CNNs) we train the system on previously labeled COAT image data. We train four different models based on the MobileNet architecture. The models vary in number of weights, and input image resolution. Results show that we can automatically classify images on a small computer like the Raspberry Pi, with an accuracy of 81.1% at 1.17 FPS, and a model size of 17Mb. In comparison a GPU computer achieves the same accuracy and model size, but it has a classification speed of 12.5 FPS.